Launched this week
Hermit
Leave ChatGPT while keeping everything it learned about you
161 followers
Leave ChatGPT while keeping everything it learned about you
161 followers
Switching LLMs means losing years of context. Anthropic's Import Memory official prompt extracts ~40 stored facts: about 2% of what's in your conversations. Hermit processes your full ChatGPT data export and generates one structured profile per ChatGPT project and recurring theme, with temporal awareness (ACTIVE vs PAST). Ready-to-paste files for Claude Memory, Claude Projects, Gemini Gems, or any LLM. Free analytics. One-shot pricing, not a subscription. Data deleted within 24h.
Free Options
Launch Team / Built With
Hermit
Oh man I went through this exact pain last year. Had like 2 years of ChatGPT conversations and switching to Claude meant losing all that built-up context about my projects and preferences. The fact that you're generating actual Claude Memory Import files is super smart — way better than just dumping raw conversation history. How accurate is the behavioral profile it generates? Like does it actually capture working style nuances or is it more surface level stuff?
Hermit
@mihir_kanzariya Thanks, that's exactly the pain I built this for!
On accuracy: it goes way beyond surface level. The pipeline generates what I call "Key Specifics": conditional behavioral instructions extracted from your conversation patterns. Not just "user is a developer" but "when user asks for code, always include error handling and types - they'll ask for it anyway" or "user prefers direct feedback over diplomatic hedging."
On my own profile it detected 95 of these across 19 topic clusters. Some of them surprised me, patterns I wasn't even conscious of but immediately recognized as true.
That said, the quality scales with your history. 50 conversations gives you a decent sketch. 500+ is where it gets eerily accurate.
The free tier gives you the analytics and clustering so you can see the depth of the analysis before committing.
Would love to hear your take if you try it!
Recently, Openai accepted a defense deal with US DOD, the same deal which Anthropic refused bc it didn't want their ai to train for auto-"killings". Seeing this, many openai users wanted to migrate away from it due to safety concerns. Before, there was no easy way to save the ai history, settings and info. Now Hermit just allowed the trasnfer to be easy.
Weird question but did you build Hermit with that in mind? @jkjakubowski
Hermit
@peterz_shu In fact: not really . I built Hermit for myself first. I'd quit ChatGPT for Claude after 3 years and 1,258 conversations, and I wanted to understand what was actually in those conversations, not just migrate them. The analysis part came first: what did I talk about, how did my usage evolve, what patterns emerged.
Then when I saw the output, I realized this could be turned into portable profiles that any LLM could use. The QuitGPT wave and the Pentagon deal happened in parallel: I didn't plan around it, but the timing made the migration angle very concrete for a lot of people.
So the honest origin is curiosity about my own data, not the news cycle. The news cycle just made it urgent for everyone else too.
Very interesting and I was actually dealing with sorting out my own transition to Claude when this showed up on PH!
I see there's currently a 100MB limit on processing exports - how would you manage larger exports? I'd assume heavy ChatGPT users would have much larger export files (I do for one!)
Hermit
@wouldntyouliketoknow Great timing then! I actually just pushed an update: limit is now 150MB (was 100MB), and error handling for large files is improved: if parsing fails you can retry without re-uploading. Most exports fall well under that, but heavy users were hitting the wall. How big is yours? If it's above 150MB I want to know, it'll help me prioritize the next bump.
@jkjakubowski Amazing!
I'm at 197MB unfortunately - I now realize that I probably did use ChatGPT a lot. Haha!
I wondered if I can minify the json to get it within 150MB, which it does, but for some reason fails hermit validation and throws an error 'doesn't look like a chatgpt export'.
Hermit
@wouldntyouliketoknow Just raised the limit to 250Mb: you should be good now! As for why minifying didn't work: Hermit validates the JSON structure by checking the first conversation object against a schema. Most JSON minifiers strip whitespace, but some also remove keys with null values (like title: null). If your minifier dropped the title or mapping fields, Hermit would reject it as "not a ChatGPT export." The raw file should work now without any minification. Let me know if you still struggle with it, happy to help you out! And it's a great feedback to make it work for everyone
This fills a real gap — people have a ton of implicit context locked in ChatGPT and there's no clean way to own it. Curious how Hermit handles storage of the exported data on your end: is it processed server-side or does everything stay local? That distinction matters a lot for the privacy-conscious users you're going after.
Hermit
@avinash_matrixgard Great question. The export is processed server-side: it has to be, since the pipeline runs through three Anthropic models (Haiku for scoring, Sonnet for clustering, Opus for synthesis). But a few things on the privacy side:
- Your data is auto-deleted within 24 hours of processing
- We use Anthropic's paid API: your conversations are never used for model training
- No human ever sees your data: the pipeline is fully automated
- The only personal info we collect is your email (for delivering results)
Fully local processing would be ideal but isn't realistic for this kind of multi-model pipeline: the export files can be up to 100MB and require several passes. The tradeoff is server-side processing with aggressive deletion. Happy to go deeper on any of this.
@jkjakubowski That's a clean privacy architecture given the constraints 24-hour deletion with no training data use is the right call for something handling personal AI exports, and being upfront about the tradeoff is the kind of thing that builds trust faster than vague privacy claims. The three-model pipeline is interesting too; Haiku for scoring then Opus for synthesis is a smart way to keep costs manageable without compromising output quality on the final layer. Curious whether you're batching the export chunks across the three passes or running them sequentially per chunk?
Hermit
@avinash_matrixgard Thanks! Sequential across the three model layers, but parallelized within each layer. Haiku scores conversations in batches of ~100 in parallel, Sonnet clusters and summarizes once all scores are in, then Opus fires once at the end on the full picture for the global synthesis. Each phase's output feeds the next, so the order is strict. But within each phase we batch aggressively. ~1000 conversations process in a few minutes. Cost-wise it mirrors the architecture: Haiku handles the bulk, Opus only fires once or twice.
@jkjakubowski That's a well-designed pipeline using Haiku for the high-volume scoring pass keeps costs linear while Opus firing once on the full picture is exactly the right place to spend the expensive tokens. The strict phase ordering with parallel batching within each phase is also how you'd want to build it for reliability; sequential phases mean a failure is isolatable and retryable without rerunning the whole job. Curious how you handle partial failures if Sonnet's clustering pass errors mid-batch, do you checkpoint or restart from Haiku?
You plan to make the Hermit a provider agnostic product? Also make it usable for anthropic and google?
Hermit
@elia_yakin hey! The output is already provider agnostic - Hermit generates .md files you can paste into Claude (Projects + Memory Import), Gemini Gems, ChatGPT Custom Instructions, or any LLM that accepts system prompts.
For input, we currently support ChatGPT exports (the biggest migration wave right now). Anthropic and Google export support is on the roadmap - if there's demand I'll prioritize it. Thanks for asking!
Notice that this product could be useful for cases of moving from one account to another within the same provider
Hermit
@elia_yakin Exactly right! That's a use case I hadn't highlighted enough. If you're moving from a personal to a work account (or merging two accounts), you lose all your context the same way. Hermit solves that too: the profiles are just portable files, doesn't matter where they end up. Good call, adding this to the landing page.